| An organic solar cell(OSC)is one kind of novel energy conversion devices.As compared with the inorganic counterpart,the OSC has the advantages of easy synthetization,low cost,flexibility and so on.However,the low power conversion efficiency(PCE)of OSC limits its applications.The design of novel materials to increase the efficiency is thus one of the key issues that needs to be addressed.The OSCs consist of the electron donor and acceptor materials.The donors are organic polymer materials,and the acceptors are often fullerene-based molecules.Recently,it has been found that non-fullerene organic molecules can significantly improve the PCE of OSCs because of their tunable light adsorption range and orbital energy levels.The emerging non-fullerene acceptors(NFAs)provide a huge chemical space for designing high-performance OSCs.At the same time,it also puts forward a scientific problem on how to find suitable acceptors in a large number of organic molecules.In this thesis,deep learning models have been constructed to address the problem.The models are trained with the dataset of NFAs including their corresponding frontier orbital energies and reorganization energies.With the trained models,the promising NFAs can be designed and screened out in a short time.The major contents and the results of this thesis are listed as follows.1.The SMILES-based and the graph-based property prediction models are built for predicting molecular properties.The architectures for the two models are convolutional neural networks(CNNs)and graph neural networks,respectively.Attention mechanisms are used in the readout parts of the prediction models in order to provide interpretability.The models are trained with the library of NFAs and their corresponding frontier orbital energies.The results show that both of the models can achieve good performance for predicting molecular frontier orbital energies,and the graph-based model performs better than the SMILES-based model.It indicates that the molecular representation based on chemical intuition has a better predictive performance.In addition,the attention mechanism incorporated in the models is good for revealing the underlying structure-property relationships of molecules.2.With the same dataset of NFAs and the corresponding SMILES strings as molecular representations,the causal and dilated CNNs are firstly used as the main architecture for molecular generation.The results show that the well-trained model can automatically generate molecules,and the frontier molecular orbital energies of the generated new molecules can be predicted precisely by the prediction model.The generated molecules have the similar property distribution as the molecules in the dataset.But the molecules generated by CNNs with different layers have the difference in diversity compared with the original dataset.In this way,one can control the diversity of generated molecules.3.In the generative model,the frontier molecular orbital energies are introduced as a part of the input to construct the conditional generation model.The results indicate that the well-trained model can generate the NFAs with given frontier molecular orbital energies.Combined with the property prediction model,it is demonstrated that the generated new molecules can expand the original chemical space.Additionally,the randomized SMILES strings are used for retraining the generative model.The retrained model can generate the required NFAs with the given molecular fragments and the targeted frontier molecular orbital energies.4.Based on the self-built small dataset of NFAs with reorganization energies,multiple machine learning models are built with various molecular representations for property prediction.Then,the models are aggregated to improve the predictive accuracy of molecular reorganization energies.Combining the molecular generation model and property prediction models,the molecules in dataset and the generated new molecules with low reorganization energies are filtered out.The molecules screened out are used for finetuning the conditional generation model.This process is repeated dozens of times.Then,the molecular conditional generation model can generate NFAs with low reorganization energies for targeted frontier orbital energies. |
